Rhythm before I discovered beatboxing
My musical interests have been broad: techno, dubstep, drum and bass, house and classical music. I was always drawn to pretending to play the drums. Then I saw beatboxing for the first time, and discovered a way to create rhythms and sounds with my own voice.
14 years of beatboxing, an RC-505, and making music on a PC
I have been beatboxing for 14 years. As a teenager, I practiced intensively: returning to new skills, refining them and trying further techniques. Rhythm, voice and experimenting with sound became a natural part of making things for me.
I later moved on to experiments with an RC-505, followed by making music on a PC. Each stage brought me closer to creating my own music and finding my own way of working with sound.
Beatboxing as input to my model-based workflow
Over the last 2 years, I have become deeply interested in making music with AI, and more recently in working with transformer models. Beatboxing is the main source material in my workflow: I start there, then develop the idea by combining models, personal techniques and further experiments with sound.
This is a creative project in its own right, combining musical curiosity with an interest in how models work. I publish the results on MistrzdYoda on YouTube.
Ideas that arrive during training
Hybrid training creates space for thinking. Musical ideas often come to me while running or during other training sessions. I return to them afterwards and try to give them a form.
Sport, technology and music meet through this process: an idea arrives while moving, and working on it leads to experiments with sound and models.
Different models, personal techniques and hybrids
I work with models including HeartMuLa and ACE-Step. I use different models and my own techniques, and create personal hybrid approaches based on these projects. I explore ways to combine their capabilities while making music.
Both the result and the process interest me: what a different model changes, how text influences sound, and which combinations help develop a musical idea.
Learning about multimodal models
Making music helps me study the architecture and operation of multimodal transformer models. I am learning how text and audio are represented, how generation is conditioned, and how different parts of a system work together to create music.
HeartMuLa describes music generation conditioned on lyrics and tags, with a separate codec handling the audio representation. ACE-Step combines diffusion-based generation, compressed audio representations and a transformer. These different architectures give me concrete reference points for learning and personal experiments.
Where to listen
I publish music at youtube.com/@MistrzdYoda. The model authors document the tools I use: HeartMuLa / heartlib and ACE-Step.